440 likes | 659 Views
Regression Wisdom. Chapter 9. No regression analysis is complete without a display of the residuals to check the linear model is reasonable.
E N D
Regression Wisdom Chapter 9
No regression analysis is complete without a display of the residuals to check the linear model is reasonable. • The residuals are what is “left over” after the model describes the relationship, the often reveal subtleties that were not clear from a plot of the original data. • Sometimes they reveal violations of the regression conditions that require our attention. Sifting Residuals for Groups
It is a good idea to look at both a histogram of the residuals and a scatterplot of the residuals vs. predicted values in the regression predicting Calories from Sugar content in cereals: • The small modes in the histogram are marked with different colors and symbols in the residual plot above. What do you see? Sifting Residuals for Groups
Two groups stand away from the central pattern. • The high residual group is “healthy” cereals that often contain more fat and therefore more calories than might be expected from looking at sugar content alone. • The low residual group is have fewer calories than what is expected from sugar content. • These observations may not lead us to a linear model but do help us understand that other factors may be a part of the story. Sifting Residuals for Groups
Examining the residuals often lead us to discover groups of observations that are different from the rest. • We may decide to analyze the groups separately using a different model for each group. • Or me may stay with the original model and note that there are groups that are a little different. • Either way the model will be wrong, but useful, so it will improve our understanding of the data. Sifting Residuals for Groups
Here’s an important unstated condition for fitting models: All the data must come from the same group. When we discover that there is more than one group in a regression, neither modeling the groups together nor modeling them apart is necessarily correct. You must determine what makes the most sense. In the following example, we see that modeling them apart makes sense. Subsets
The figure shows regression lines fit to calories and sugar for each of the three cereal shelves in a supermarket: Subsets
Linear regression only works for linear models. (That sounds obvious, but when you fit a regression, you can’t take it for granted.) A curved relationship between two variables might not be apparent when looking at a scatterplot alone, but will be more obvious in a plot of the residuals. Remember, we want to see “nothing” in a plot of the residuals. Getting the “Bends”
The scatterplot of residuals against Duration of emperor penguin dives holds a surprise. The Linearity Assumption says we should not see a pattern, but instead there is a bend. Even though it means checking the Straight Enough Condition after you find the regression, it’s always good to check your scatterplot of the residuals for bends that you might have overlooked in the original scatterplot. Getting the “Bends” (cont.)
Linear models give a predicted value for each case in the data. We cannot assume that a linear relationship in the data exists beyond the range of the data. The farther the new x value is from the mean in x, the less trust we should place in the predicted value. Once we venture into new x territory, such a prediction is called an extrapolation. Extrapolation: Reaching Beyond the Data
Extrapolations are dubious because they require the additional—and very questionable — assumption that nothing about the relationship between x and y changes even at extreme values of x. Extrapolations can get you into deep trouble. You’re better off not making extrapolations. Extrapolation
The U.S. Census reports the median age at first marriage for men and women. Here’s a regression of age at first marriage for men against year at every census from 1890 to 1940: • R-squared = 92.6% • s = 0.2417 Example
The regression line is • The slope of -0.04 years of age per year of century shows that the first marriage age was falling at a rate of about 4 years per century. Men were getting married for the first time at a younger and younger age. • The intercept can be interpreted as the predicted mean age of men at first marriage in 1890 (counted as year ). The high suggests a linear pattern. Can we extrapolate? Example
Use the linear model to predict the average age of men at first marriage in the year 2000. • In fact, by the year 2000, the median age at first marriage for men was almost 27 years. What went wrong? Example
Extrapolation is always dangerous. But, when the x-variable in the model is time, extrapolation becomes an attempt to peer into the future. Knowing that extrapolation is dangerous doesn’t stop people. The temptation to see into the future is hard to resist. Here’s some more realistic advice: If you must extrapolate into the future, at least don’t believe that the prediction will come true. Predicting the Future
Outlying points can strongly influence a regression. Even a single point far from the body of the data can dominate the analysis. Any point that stands away from the others can be called an outlier and deserves your special attention. The outcome of the 2000 U.S. presidential election was determined in Florida amid much controversy. The main race was between George W. Bush and Al Gore, but two minor candidates played a significant role. Outliers, Leverage, and Influence
To the political right of the main party candidates was Pat Buchanan, while to the political left of the main party candidates was Ralph Nader. • Generally, Nader earned more votes than Buchanan throughout the state. • We would expect counties with larger vote totals to give more votes to each candidate. Outliers, Leverage, and Influence
Here’s a regression relating Buchanan’s vote totals by county in the state of Florida to Nader’s: • Dependent variable: Buchanan vote • R-squared: 42.8% • Regression model: Outliers, Leverage, and Influence
The regression model says that in each county, Buchanan received about 0.14 time the vote Nader received, starting from a base of 50.3 votes. • It seems reasonable with an of almost 43%, however…. • We violated all 3 laws of Data Analysis by not making a picture.
The “butterfly” ballot used only in Palm Beach county was a source of much controversy. • Some claim the format of the ballot confused voters so that some who intended to vote for the Democrat Al Gore punched the wrong hole next to his name and actually voted for Pat Buchanan.
When Palm Beach county is removed from the regression, jumps from 42.8% to 82.1% and the slope of the line changes to 0.1. • This suggests that Buchanan only received about 10% of the vote that Nader did. • With more than 82% of the variance of the Buchanan vote accounted for, the model with Palm Beach county omitted fits better. • The linear model doesn’t fit points with large residuals very well. • Because they seem to be different from the other cases, it is important to pay special attention to points with large residuals.
A data point can also be unusual if its x-value is far from the mean of the x-values. Such points are said to have high leverage. Outliers, Leverage, and Influence
A point with high leverage has the potential to change the regression line. We say that a point is influential if omitting it from the analysis gives a very different model. Influence depends on both leverage and residual. A case with high leverage that sits right on the line fit to the rest of the data is not influential. Removing that case won’t change the slope even if it does affect Outliers, Leverage, and Influence
A case with modest leverage but a very large residual (Palm Beach county) can be influential. • If a point has enough leverage, it can pull the line right to it. Then it’s highly influential but its residual is small. • The only way to be sure is to fit both regressions. Outliers, Leverage, and Influence
A point with so much influence that it pulls the regression line close to it can make its residual deceptively small. • The extraordinarily large shoe size gives the data point high leverage. Wherever the IQ is, the line will follow!
When we investigate an unusual point, we often learn more about the situation than we could have learned from the model alone. You cannot simply delete unusual points from the data. You can, however, fit a model with and without these points as long as you examine and discuss the two regression models to understand how they differ. Outliers, Leverage, and Influence
Warning: Influential points can hide in plots of residuals. Points with high leverage pull the line close to them, so they often have small residuals. You’ll see influential points more easily in scatterplots of the original data or by finding a regression model with and without the points. Outliers, Leverage, and Influence (cont.)
Turn to page 206 and do the just checking exercises. • 1. Not high leverage, not influential, large residual • 2. High leverage, not influential, small residual • 3. High leverage, influential, not large residual Checkpoint
No matter how strong the association, no matter how large the R2value, no matter how straight the line, there is no way to conclude from a regression alone that one variable causes the other. There’s always the possibility that some third variable is driving both of the variables you have observed. With observational data, as opposed to data from a designed experiment, there is no way to be sure that a lurking variable is not the cause of any apparent association. Lurking Variables and Causation
The following scatterplot shows that the average life expectancy for a country is related to the number of doctors per person in that country: Lurking Variables and Causation Why square root?
The strong association, , seems to confirm the expectation the more doctors per person improves healthcare, leading to longer lifetimes and a larger life expectancy. • The strength of the association would seem to argue that we should send more doctors to developing countries to increase life expectancy. • That conclusion is about the consequences of a change. If we increase the number of doctors, should we predict the life expectancy will increase? • This is a casual explanation arguing that adding doctors causes greater life expectancy.
This new scatterplot shows that the average life expectancy for a country is related to the number of televisions per person in that country: Lurking Variables and Causation
The positive association of this scatterplot is even stronger than it was in the doctors/life expectancy plot • Should we conclude that increasing the number of televisions will extend lifetimes?? • Since televisions are cheaper than doctors, send TVs to countries with low life expectancies in order to extend lifetimes. Right? • How about considering a lurking variable? That makes more sense… • Countries with higher standards of living have both longer life expectancies and more doctors (and TVs!). • If higher living standards cause changes in these other variables, improving living standards might be expected to prolong lives and increase the numbers of doctors and TVs.
Scatterplots of statistics summarized over groups tend to show less variability than we would see if we measured the same variable on individuals. This is because the summary statistics themselves vary less than the data on the individuals do. Working With Summary Values
In chapter 7, we looked at the heights and weights of students. The correlation was 0.644 so There is a strong, positive, linear association between weight (in pounds) and height (in inches) for men:
If instead of data on individuals we only had the mean weight for each height value, we would see an even stronger association:
Means vary less than individual values. Scatterplots of summary statistics show less scatter than the baseline data on individuals. This can give a false impression of how well a line summarizes the data. There is no simple correction for this phenomenon. Once we have summary data, there’s no simple way to get the original values back. Working With Summary Values
Make sure the relationship is straight. Check the Straight Enough Condition. Be on guard for different groups in your regression. If you find subsets that behave differently, consider fitting a different linear model to each subset. Beware of extrapolating. Beware especially of extrapolating into the future! What Can Go Wrong?
Look for unusual points. Unusual points always deserve attention and may well reveal more about your data than the rest of the points combined. Beware of high leverage points, and especially those that are influential. Such points can alter the regression model a great deal. Consider comparing two regressions. Run regressions with extraordinary points and without and then compare the results. What Can Go Wrong?
Treat unusual points honestly. Don’t just remove unusual points to get a model that fits better. Beware of lurking variables—and don’t assume that association is causation. Watch out when dealing with data that are summaries. Summary data tend to inflate the impression of the strength of a relationship. What Can Go Wrong?
There are many ways in which a data set may be unsuitable for a regression analysis: Watch out for subsets in the data. Examine the residuals to re-check the Straight Enough Condition. The Outlier Condition means two things: Points with large residuals or high leverage (especially both) can influence the regression model significantly. Perform regression analysis with and without such points to see their impact. What have we learned?
Even a good regression doesn’t mean we should believe the model completely: Extrapolation far from the mean can lead to silly and useless predictions. An R2 value near 100% doesn’t indicate that there is a causal relationship between x and y. Watch out for lurking variables. Watch out for regressions based on summaries of the data sets. These regressions tend to look stronger than the regression on the original data. What have we learned?